SECAI Core · phase 11 of 15
Data poisoning
When misinformation or malicious data is used in the training dataset, causing bias and changes in behavior
The Explain card
- Plain English
- Data poisoning plants misinformation or malicious samples in a training dataset so the resulting model learns bias, wrong answers, or hidden behaviors.
- Example
- An attacker edits a few hundred pages on a public wiki that a company scrapes for training. The model later confidently repeats the attacker's false claims about a competitor's product.
- Why it matters
- Training data is an attack surface that is huge, hard to audit, and often scraped from the open internet. Defenders track data provenance, filter sources, and test models for unexpected behaviors.
- Hook
- Poison the well and everyone downstream drinks it.
Where it sits in the deck
Phase 11: Threat Landscape: Attack Vectors and Adversarial Techniques
With defences named, learn what they must defend against — the full catalogue of attack techniques targeting AI systems, their inputs, outputs, training pipelines, and supply chains.
